DeepSeek has open-sourced a complete programming toolkit for Huawei’s Ascend AI accelerators, marking one of the most concrete attempts yet to build a domestic alternative to Nvidia’s entrenched CUDA ecosystem. The announcement, made through the company’s social channels, positions the release as a foundational step toward an independent GPU software stack that can run large AI workloads without relying on foreign platforms.
At the core of the effort is TileLang, a high-level programming language crafted to be universal and approachable while still squeezing maximum performance from the hardware. Alongside it, DeepSeek is releasing compute and communication libraries such as DeepGEMM, FlashMLA, TileKernel, DeepSelect, and DeepEP, all tuned for Huawei’s latest Ascend 950 chips and capable of scaling across supernode configurations that link up to 128 accelerators for heavy training jobs.
The initiative arrives as Chinese technology companies accelerate their shift away from Nvidia hardware under the pressure of U.S. export controls on advanced semiconductors. DeepSeek, which once leaned heavily on Nvidia’s H800 GPUs, has been adapting its models to run on domestic silicon and is reportedly planning to deploy at least 160,000 Huawei Ascend accelerators in a new data center in Inner Mongolia, underscoring the strategic stakes behind the software push.
Still, the real test lies in whether this toolkit can attract enough developers to form a viable ecosystem around Huawei’s chips. CUDA’s grip on the industry comes less from raw technical advantage than from years of accumulated libraries, tools, and community knowledge. By making its stack free and designing TileLang to feel familiar to CUDA programmers, DeepSeek is trying to lower the barrier to entry, but winning over the broader AI engineering community will determine whether this becomes a true rival or a niche alternative.
#AI #DeepSeek #Huawei #Ascend #CUDA #TileLang #Semiconductors #ChinaTech #OpenSource #GPU